# AI Hardware Refresh Planner: When Does Upgrading Your GPU Fleet Pay Off

> Source: https://netray.co/tools/ai-hardware-refresh-planner
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Weigh your current GPU fleet's remaining book value against the cost of refreshing to a newer generation, factoring in performance-per-watt gains and power savings.

## Category

on-prem-ai

## Tool

**Kind:** calculator

**Inputs:**

#### Current GPU count

**Key:** current_gpu_count

**Type:** number

**Default:** 32

**Min:** 1

**Max:** 2048

**Step:** 1

**Unit:** GPUs

**Help:** Accelerators in your existing fleet being considered for refresh.

#### Age of current fleet

**Key:** current_gpu_age_years

**Type:** number

**Default:** 3

**Min:** 0

**Max:** 10

**Step:** 0.5

**Unit:** years

**Help:** Time since the current GPUs were purchased and put into service.

#### Original purchase price per GPU

**Key:** current_gpu_purchase_price

**Type:** number

**Default:** 28000

**Min:** 500

**Max:** 100000

**Step:** 500

**Unit:** $/GPU

**Help:** What you originally paid per accelerator when the current fleet was purchased.

#### Annual depreciation rate

**Key:** annual_depreciation_pct

**Type:** slider

**Default:** 20

**Min:** 10

**Max:** 35

**Step:** 5

**Unit:** %

**Help:** Straight-line depreciation rate used to estimate remaining book value.

#### New generation GPU price

**Key:** new_gpu_price

**Type:** number

**Default:** 45000

**Min:** 1000

**Max:** 100000

**Step:** 500

**Unit:** $/GPU

**Help:** Purchase price per unit for the generation you would refresh to.

#### Performance-per-watt improvement

**Key:** perf_per_watt_gain_pct

**Type:** slider

**Default:** 40

**Min:** 5

**Max:** 100

**Step:** 5

**Unit:** %

**Help:** How much more useful work the new generation does per watt versus your current fleet, from vendor benchmarks or your own testing.

#### Current fleet annual power and cooling cost

**Key:** current_annual_power_cost

**Type:** number

**Default:** 45000

**Min:** 1000

**Max:** 2000000

**Step:** 1000

**Unit:** $/year

**Help:** Total electricity and cooling cost for the current fleet at its typical utilization.

**Outputs:**

#### Remaining book value of current fleet

**Key:** current_book_value

**Expr:** current_gpu_count * current_gpu_purchase_price * max(0, 1 - annual_depreciation_pct / 100 * current_gpu_age_years)

**Format:** currency

**Description:** What the existing fleet is still worth on paper after straight-line depreciation.

**Primary:** true

#### Capex to refresh to the new generation

**Key:** refresh_capex

**Expr:** current_gpu_count * new_gpu_price

**Format:** currency

**Description:** Cost to replace the current fleet one-for-one with the new generation.

**Primary:** true

#### Net cost of refreshing

**Key:** net_refresh_cost

**Expr:** refresh_capex - current_book_value

**Format:** currency

**Description:** True incremental cost of the refresh after accounting for the current fleet's remaining book value.

#### Projected power cost after refresh

**Key:** projected_new_power_cost

**Expr:** current_annual_power_cost / (1 + perf_per_watt_gain_pct / 100)

**Format:** currency

**Description:** Estimated annual power cost to deliver the same useful work on the more efficient new generation.

#### Annual power savings from refreshing

**Key:** annual_power_savings

**Expr:** current_annual_power_cost - projected_new_power_cost

**Format:** currency

**Description:** Yearly reduction in power and cooling spend attributable to the performance-per-watt gain alone.

#### Payback period from power savings alone

**Key:** simple_payback_years

**Expr:** net_refresh_cost / max(annual_power_savings, 1)

**Format:** years

**Description:** Years for power savings alone to offset the net refresh cost. Compute throughput gains almost always dominate this decision.

**Disclaimer:** Planning estimate based on straight-line depreciation and vendor-reported performance-per-watt figures. Actual refresh decisions should also weigh compute throughput gains, new capability requirements, and secondary market resale value of retired hardware.

## Content

**Intro:** This free AI hardware refresh planner models whether and when to upgrade your GPU fleet to a newer generation, and it is built for infrastructure leads planning multi-year capital cycles for on-prem AI. Enter your current fleet's age, original cost, and power draw alongside the new generation's price and performance-per-watt improvement, and the tool returns remaining book value, net refresh cost, and the payback period from power savings alone. The honest answer for most fleets is that power savings alone rarely justify an early refresh, throughput and capability gains almost always carry the real business case.

**Sections:**

#### Why power savings alone rarely justify a refresh

**Body:** A three-year-old fleet of 32 H100s originally purchased at $28,000 each, depreciated at 20% per year, retains roughly $269,000 of book value. Refreshing to B200 at $52,000 per GPU costs $1,664,000 in new capex, for a net refresh cost near $1,395,000 after netting out the retained book value. Even a generous 40% performance-per-watt gain on a $45,000 annual power bill saves only about $12,900 per year, a payback period well over a century on power alone. This is not an argument against refreshing, it is a reminder that the real justification has to come from throughput, new model capability, or capacity headroom, not electricity savings.

#### What actually drives a sound refresh decision

**Body:** Performance-per-watt is real and worth tracking, but the business case for a GPU refresh almost always rests on compute throughput per dollar, memory capacity for larger models, and whether the current fleet can even run the models your roadmap requires. A newer generation delivering meaningfully more tokens per second per GPU can reduce total GPU count needed for the same workload, which changes the capex comparison far more than power savings ever will.

**Bullets:**

- Compare tokens-per-second-per-dollar across generations for your actual target models, not vendor marketing benchmarks.
- Check whether your current fleet's VRAM can even fit the model sizes your roadmap requires before evaluating refresh economics.
- Factor in secondary market resale value for the retired fleet, which can materially offset net refresh cost.
- Weigh warranty expiration and support contract renewal costs, which often rise sharply as hardware ages past three years.

#### Reading remaining book value and net cost

**Body:** Remaining book value tells you what the current fleet is theoretically worth on paper, useful for internal capital planning conversations even though it does not reflect real resale value. Net refresh cost, the new capex minus that book value, is the number that should anchor your business case discussion, since it represents the true incremental capital commitment rather than the full sticker price of the new hardware.

#### How Netray plans hardware refresh cycles

**Body:** Netray helps manufacturers plan multi-year GPU refresh cycles that align hardware capability with actual model and workload roadmaps rather than refreshing on a fixed calendar schedule regardless of need. We benchmark real throughput gains for your specific workloads across generations before recommending a refresh, and coordinate resale or repurposing of retired hardware where it still has useful life for lighter workloads. Engagements typically start with a fleet assessment that produces a multi-year refresh roadmap.

**Cta:** Get a benchmarked refresh roadmap that weighs throughput, capability, and cost across your actual GPU generations.

**Faq:**

#### Should we refresh on a fixed schedule or based on need?

**Answer:** Based on need in almost every case. A fixed three-year refresh schedule regardless of workload requirements often wastes capital on hardware that still meets your needs, while a workload-driven approach refreshes when the current fleet genuinely cannot deliver required throughput, capacity, or model support. Use warranty expiration as a secondary trigger to evaluate, not an automatic refresh decision.

#### What can we do with retired GPU hardware?

**Answer:** Common options include reselling on the active secondary market, which retains meaningful value for hardware within two generations of current, repurposing for lighter internal workloads such as development, testing, or smaller model inference, or donating to research partners for tax and goodwill benefit. Factor whichever path you plan into your net refresh cost calculation.

#### How much does performance-per-watt actually improve between GPU generations?

**Answer:** It varies by generation and workload, but recent NVIDIA generations have delivered performance-per-watt gains in the range of 30-70% for inference workloads, driven by architectural improvements and memory bandwidth increases rather than raw power draw increases alone. Validate vendor-published figures against your actual workload before using them for a business case, since gains vary significantly by model type and batch size.

#### Is it ever worth refreshing before the current fleet's warranty expires?

**Answer:** Sometimes, if a new generation unlocks a model size or throughput requirement your roadmap genuinely needs and the current fleet cannot deliver, or if a major workload expansion is planned that would otherwise require buying additional units of the older generation. Absent a clear capability or capacity gap, waiting until warranty expiration to evaluate a refresh is usually the more capital-efficient approach.

## Search terms

GPU hardware refresh calculator, AI infrastructure depreciation, GPU generation upgrade payback, performance per watt GPU, data center GPU refresh cycle, GPU fleet upgrade planning, H100 to B200 upgrade cost

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json